BUDGETS AND RESOURCE LIMITS

EXAMPLES Example 1: A token budget A run is given 500,000 tokens in total. After each step the harness subtracts the tokens used, so it always knows how much fuel remains. At 400,000 tokens used it logs a warning, and at 480,000 it tells the model to wrap up. Example 2: Per-workspace tracking A company gives its marketing workspace a shared limit of 200 dollars per day. Ten agents run there, and each also has a personal limit of 30 dollars. One agent develops a bug and begins to loop, but it stops at its own 30 dollars, and the workspace total never approaches the shared cap. Example 3: Graceful termination A research agent hits its time budget while a search call is in progress. The harness waits for that call to finish, refuses to start another, and uses its reserved tokens to produce a note that says which sources were read and which were not. The user receives a partial but useful answer instead of an error message. Example 4: Bounded retries A tool call fails because a server is busy. The harness retries after one second, then two, then four, adding a small random delay each time. After three failed retries it tells the agent that the tool is unavailable, and the agent either tries a different tool or reports the problem. PRACTICE LAB CONNECTION Add the budget tracker to the Resume Review Agent lab from Chapter 65. Then deliberately trigger the runaway-loop bug from Chapter 68 and test the prompts below against what happens.